用於磁振造影 到電腦斷層 影像轉換的深度學習框架

Abstract

磁振造影（MRI）可在無游離輻射的狀況下提供優異的軟組織對比，但缺乏進行劑量計算所需的電子密度資訊. 因此，臨床流程必須同時依賴 MRI 與電腦斷層掃描（CT），增加流程複雜度並帶來配準誤差. 由 MRI 生成合成 CT（sCT）可望實現僅使用 MRI 的放射治療計畫，但因 MRI 與 CT 強度之間呈現非線性對應而具高度挑戰性. 本研究提出 Med2Transformer，一種 3D 雙分支編碼器模型，用於將 MRI 轉換為 sCT. 該架構結合卷積式與 Transformer 編碼器，並採用多尺度移窗自注意力機制，同時捕捉細微解剖結構及更廣泛的上下文資訊. 模型透過組合體素重建, 對抗及感知損失的複合損失函數進行訓練，以提升解剖準確性與影像強度一致性. Med2Transformer 於涵蓋腦部, 骨盆和頭頸部之公有與私有資料集上進行評估，在所有解剖區域均達到最先進表現；其中於頭部區域獲得平均絕對誤差（MAE）74.58 HU, 結構相似度指數（SSIM）0.8639 及峰值訊雜比（PSNR）27.73 dB. 幾何一致性評估亦顯示較高的 Dice 係數及較低的 Hausdorff95 距離，證實解剖保真度. 此外，單病例 CyberKnife 劑量學評估顯示劑量分布符合臨床可接受標準，20 個解剖結構的平均劑量誤差為 3.83%. 研究結果顯示，Med2Transformer 能生成精確且具泛化能力的 sCT 影像，支援 MRI 單影像放射治療計畫，並提供可擴充的臨床整合解決方案.

Magnetic Resonance Imaging (MRI) provides high soft-tissue contrast without ionizing radiation but lacks electron density information needed for dose calculation. As a result, clinical workflows depend on both MRI and CT, increasing complexity and registration errors. Synthetic CT (sCT) generation from MRI enables MRI-only radiotherapy planning but remains challenging due to the non-linear mapping between MRI and CT intensities. This study proposes Med2Transformer, a 3D dual-branch encoder model for MRI-to-synthetic CT generation. The architecture merges convolutional and transformer-based encoders, employing multi-scale shifted-window self-attention to effectively represent fine-grained anatomical structures along with broader contextual patterns. The model is trained using a composite loss function comprising voxel-wise reconstruction, adversarial, and perceptual losses to enhance anatomical accuracy and intensity consistency. Med2Transformer was evaluated on public and private datasets spanning brain, pelvis, and head-and-neck regions. It achieved state-of-the-art performance across all anatomical sites, with a mean absolute error (MAE) of 74.58 HU, structural similarity index (SSIM) of 0.8639, and peak signal-to-noise ratio (PSNR) of 27.73 dB in the head region. Geometric consistency assessments further confirmed anatomical fidelity, as reflected by higher Dice coefficients and lower Hausdorff95 distances. Additionally, a single-case CyberKnife dosimetric evaluation demonstrated clinically acceptable dose distributions, with an average mean dose error of 3.83% across 20 anatomical structures. These findings demonstrate that Med2Transformer generates accurate and generalizable sCT images, supporting MRI-only radiotherapy planning and offering a scalable solution for clinical integration.